Skip to main content
This page is the engineering ledger of where Nelo actually is: what has run, what is only written, and what is deliberately stubbed. Investors, partners and new engineers should read it before anything that sounds like a claim.
Three states are kept apart, because conflating them is how a payments company ends up claiming something untrue:
  • Proven — a test that has actually executed, or a run on devnet
  • Written — code exists and compiles, but has never run
  • Declared stub — deliberately not real, and labelled as such at every call site

Where it stands

The trust model

Proven, on devnet. Vault, voucher redemption with the signature verified by the precompile on a real validator, 128-slot replay window, timelocked withdrawal, conflict freeze, double-spend refused.Caveat: run with a software P-256 key, not StrongBox.

The ordinary sale

Runs on a handset, in signed release builds. The till charges in naira, Solflare pays its Solana Pay code, and the balance reads the chain.The offline sale also runs on two phones (StrongBox payer, vault funded): scanned both ways and recorded, and settled through the relayer once its fee payer is funded on devnet; see Testing.

Trust Stake

Proven, on devnet. 17 curve tests off chain, ~19 on chain in LiteSVM, and RiskConfig now live on a real validator with redeem_voucher exercised against its new account struct.

Settlement

Ledger proven: double-entry, idempotent, reversible, reconcilable. Cash-out live on mainnet: merchant USDC paid out in naira to a Nigerian bank through paj.cash’s production v2 API, with signed webhooks.

By component

Next

1

Launch

Play Store and Solana dApp Store listings, UseAzza live alongside paj.cash, and the first merchant cluster in Lagos. See the roadmap.
2

Hosted infrastructure

The relayer and settlement service move off a single machine onto managed hosting with a fixed domain, secrets in a manager and alerting on the fee payer’s SOL.
3

Audit, then offline payments on mainnet

An external review of nelo_vault, then vaults and redemptions on mainnet with low floor limits until it closes.
Also open: Privy has not enabled SMS for Nigeria on this app (a dashboard setting), so phone sign-up does not complete yet; wallet sign-up does.

Decisions waiting on a person

Pyth publishes no NGN feed — 39 FX pairs, and the naira is not among them. Of the markets the plan names, Manila (PHP) is covered; Lagos is not.Either pick a covered launch currency, or source NGN from Switchboard or a commercial feed. Hermes also requires an API key for prices. Until one is settled, the till runs a configured rate and says so on screen.
The reserve line is settled. The reserve model found the 0.20% line did not cover expected loss — 28.07 bps implied — and the rate charged is now 29 bps, which took the net take rate to 0.61%; with the payout spread now 0.80%, it is 0.91%.The premium claim is not. Reserve relief funds a premium of ~1.001×, not the illustrative 1.5×. That is a change to how the pitch describes it, not to the architecture.
The curve raises required reserve below ~$250 of staked value, and the $500 hard cap needs $81,000 of stake to bind. Neither parameter can be fixed without the other.Both are RiskConfig updates rather than redeploys.
report_conflict only freezes on two vouchers at the same sequence. Successive sequences with inconsistent remaining_after do not trigger it, and that residual is what the insurance line prices.Accepting that pair as proof would shorten the window a compromised device keeps trading in. It changes the fraud model, so it is flagged rather than taken.

What is deliberately not built

Named rather than discovered, because a named stub reads as confidence and a discovered one reads as spin:
  • Paying out of the reserve. slash moves a frozen vault’s whole stake into the reserve, an account owned by the risk-config PDA, so no key can move it. Nothing pays out of it yet. Deciding who is compensated, by how much and on whose say-so is a commercial and legal decision, and the program should not invent the answer.
  • Cash-outs in the double-entry journal. The journal was modelled on Nelo holding merchants’ money. Cash-outs are non-custodial, so only Nelo’s fee belongs in it, and that posting is not written.
  • Embedded-wallet recovery. needs-recovery is detected and reported; useRecoverEmbeddedWallet is not wired, so the retry on that screen will not clear it. It arises when a wallet has to be reconstructed on a new device, which is exactly the case a merchant who changes phones hits — so it is a gap with a date on it, not a permanent omission.
  • Moving a vault to a new phone. The program has no instruction to enrol a new device key on an existing vault, so a payer who loses the phone’s key needs a new wallet.
  • The payer chooses their own floor limit. initialize_vault accepts any floor_limit up to the platform hard cap, which makes staking pointless for anyone who asks for the cap. The payer app does not expose the knob and asks for $50, but the program should bound it.
  • NFC. QR first, as planned; NFC has not been started.
  • Holes in the replay window. The window advances only over contiguous redeemed sequences, so a voucher that is never redeemed leaves a hole the base cannot pass. After 127 more vouchers the vault can no longer pay offline. Fixing it is a program change, not yet designed.
  • Relay hosting. The relayer runs on one machine behind a fixed ngrok domain. Production needs managed hosting, secrets in a manager and monitoring; that is the next infrastructure milestone.
  • Payer onboarding without SOL. The payer app still enrols and deposits through the payer’s own wallet, which needs SOL for fees and rent. The relayer cannot sponsor those instructions as they stand, because they require the vault owner’s signature.